[PATCH v2] mm: nommu: free unused resources when mremap shrinks the vma

[PATCH v2] mm: nommu: free unused resources when mremap shrinks the vma

[Hajime Tazaki]

When shrinking a VMA via mremap, the bounds are modified directly:
mm/nommu.c:do_mremap() {
    ...
    vma->vm_end = vma->vm_start + new_len;
    ...
}
This shrink the VMA without updating its bounds in the maple tree.
If the maple tree (mm->mm_mt) still contains the old bounds, a user
process could access the freed portion. The stale maple tree would
incorrectly return the shrunk VMA for an address past its new vm_end.

This commit fixes this issue by calling vmi_shrink_vma() when shrink
happens.  Additionally, if a file-backed map is to be shrunk, it reports
-EINVAL like do_munmap() does.

The issue is reported by Sashiko review, linked below.

Link: https://sashiko.dev/#/patchset/20260702012546.665383-1-thehajime@gmail.com
Link: https://sashiko.dev/#/patchset/20260710021028.892645-1-thehajime%40gmail.com
Signed-off-by: Hajime Tazaki <thehajime@gmail.com>

--

v1 -> v2:
- handle error when vmi_shrink_vma() failed (reported by Sashiko)
- prevents mremap() with being shrunk for file-backed one like munmap()

v1: https://lore.kernel.org/linux-mm/20260710021028.892645-1-thehajime@gmail.com/
---
 mm/nommu.c | 20 +++++++++++++++++++-
 1 file changed, 19 insertions(+), 1 deletion(-)

diff --git a/mm/nommu.c b/mm/nommu.c
index 852ec9bd0505..c8effbffa0f0 100644
--- a/mm/nommu.c
+++ b/mm/nommu.c
@@ -1597,8 +1597,26 @@ static unsigned long do_mremap(unsigned long addr,
 	if (new_len > vma->vm_region->vm_end - vma->vm_region->vm_start)
 		return (unsigned long) -ENOMEM;
 
+	/* like do_munmap(), we're allowed to shrink an anonymous VMA but not a file-backed one */
+	if (vma->vm_file)
+		return (unsigned long) -EINVAL;
+
 	/* all checks complete - do it */
-	vma->vm_end = vma->vm_start + new_len;
+	if (new_len < old_len) {
+		/* shrink only happens addr + new_len and old_len are in different pages */
+		VMA_ITERATOR(vmi, current->mm, addr);
+		int ret;
+
+		/* vmi_shrink_vma() needs from/to pointers to be removed,
+		 * (mainly used in munmap) so, specify them.
+		 */
+		ret = vmi_shrink_vma(&vmi, vma, addr + new_len, addr + old_len);
+		if (ret < 0)
+			return (unsigned long) ret;
+	} else {
+		/* when there are no shrink, update vma.  */
+		vma->vm_end = vma->vm_start + new_len;
+	}
 	return vma->vm_start;
 }
 
-- 
2.43.0

Re: [PATCH v2] mm: nommu: free unused resources when mremap shrinks the vma

[Hajime Tazaki]

On Fri, 10 Jul 2026 14:46:48 +0900,
Hajime Tazaki wrote:
> 
> When shrinking a VMA via mremap, the bounds are modified directly:
> mm/nommu.c:do_mremap() {
>     ...
>     vma->vm_end = vma->vm_start + new_len;
>     ...
> }
> This shrink the VMA without updating its bounds in the maple tree.
> If the maple tree (mm->mm_mt) still contains the old bounds, a user
> process could access the freed portion. The stale maple tree would
> incorrectly return the shrunk VMA for an address past its new vm_end.
> 
> This commit fixes this issue by calling vmi_shrink_vma() when shrink
> happens.  Additionally, if a file-backed map is to be shrunk, it reports
> -EINVAL like do_munmap() does.
> 
> The issue is reported by Sashiko review, linked below.
> 
> Link: https://sashiko.dev/#/patchset/20260702012546.665383-1-thehajime@gmail.com
> Link: https://sashiko.dev/#/patchset/20260710021028.892645-1-thehajime%40gmail.com
> Signed-off-by: Hajime Tazaki <thehajime@gmail.com>

I think the original issue reported by Sashiko is addressed, but found
another existing issue, which might take a bit to look into detail
(e.g., currently under nommu we cannot map (actually kernel_read())
/dev/zero, vma_set_anonymous() isn't called for /dev/zero on nommu,
etc).

So I will address this in a different patch, not in this one.

-- Hajime

Re: [PATCH v2] mm: nommu: free unused resources when mremap shrinks the vma

[Lorenzo Stoakes]

On Fri, Jul 10, 2026 at 04:30:08PM +0900, Hajime Tazaki wrote:
>
> On Fri, 10 Jul 2026 14:46:48 +0900,
> Hajime Tazaki wrote:
> >
> > When shrinking a VMA via mremap, the bounds are modified directly:
> > mm/nommu.c:do_mremap() {
> >     ...
> >     vma->vm_end = vma->vm_start + new_len;
> >     ...
> > }
> > This shrink the VMA without updating its bounds in the maple tree.
> > If the maple tree (mm->mm_mt) still contains the old bounds, a user
> > process could access the freed portion. The stale maple tree would
> > incorrectly return the shrunk VMA for an address past its new vm_end.
> >
> > This commit fixes this issue by calling vmi_shrink_vma() when shrink
> > happens.  Additionally, if a file-backed map is to be shrunk, it reports
> > -EINVAL like do_munmap() does.
> >
> > The issue is reported by Sashiko review, linked below.
> >
> > Link: https://sashiko.dev/#/patchset/20260702012546.665383-1-thehajime@gmail.com
> > Link: https://sashiko.dev/#/patchset/20260710021028.892645-1-thehajime%40gmail.com
> > Signed-off-by: Hajime Tazaki <thehajime@gmail.com>
>
> I think the original issue reported by Sashiko is addressed, but found
> another existing issue, which might take a bit to look into detail
> (e.g., currently under nommu we cannot map (actually kernel_read())
> /dev/zero, vma_set_anonymous() isn't called for /dev/zero on nommu,
> etc).
>
> So I will address this in a different patch, not in this one.

OK thanks sorry for not getting to these quicker, between jobs atm and my inbox
is an ongoing disaster :)

>
> -- Hajime

Thanks, Lorenzo